Merged
feat(quant): dequant_bias 支持非连续 Tensor(手写 aclnn 连续化+回写,修复 #8693) #14378
feat(quant): dequant_bias 支持非连续 Tensor(手写 aclnn 连续化+回写,修复 #8693) #14378
Merged
fengjiawei1Add on 8 days ago
Total: 8 files changed+324-16
@@ -96,7 +96,7 @@ aclnnStatus aclnnDequantBias(
96 <td>INT32</td>96 <td>INT32</td>
97 <td>ND</td>97 <td>ND</td>
98 <td>2</td>98 <td>2</td>
99- <td>×</td>99+ <td>√</td>
100 </tr>100 </tr>
101 <tr>101 <tr>
102 <td>weightScale(aclTensor*)</td>102 <td>weightScale(aclTensor*)</td>
@@ -106,7 +106,7 @@ aclnnStatus aclnnDequantBias(
106 <td>FLOAT,BFLOAT16</td>106 <td>FLOAT,BFLOAT16</td>
107 <td>ND</td>107 <td>ND</td>
108 <td>1</td>108 <td>1</td>
109- <td>×</td>109+ <td>√</td>
110 </tr>110 </tr>
111 <tr>111 <tr>
112 <td>activateScaleOptional(aclTensor*)</td>112 <td>activateScaleOptional(aclTensor*)</td>
@@ -116,7 +116,7 @@ aclnnStatus aclnnDequantBias(
116 <td>FLOAT</td>116 <td>FLOAT</td>
117 <td>ND</td>117 <td>ND</td>
118 <td>1</td>118 <td>1</td>
119- <td>×</td>119+ <td>√</td>
120 </tr>120 </tr>
121 <tr>121 <tr>
122 <td>biasOptional(aclTensor*)</td>122 <td>biasOptional(aclTensor*)</td>
@@ -126,7 +126,7 @@ aclnnStatus aclnnDequantBias(
126 <td>FLOAT,BFLOAT16,FLOAT16,INT32</td>126 <td>FLOAT,BFLOAT16,FLOAT16,INT32</td>
127 <td>ND</td>127 <td>ND</td>
128 <td>1</td>128 <td>1</td>
129- <td>×</td>129+ <td>√</td>
130 </tr>130 </tr>
131 <tr>131 <tr>
132 <td>outputDtype(int64_t)</td>132 <td>outputDtype(int64_t)</td>
@@ -146,7 +146,7 @@ aclnnStatus aclnnDequantBias(
146 <td>FLOAT16、BFLOAT16</td>146 <td>FLOAT16、BFLOAT16</td>
147 <td>ND</td>147 <td>ND</td>
148 <td>2</td>148 <td>2</td>
149- <td>×</td>149+ <td>√</td>
150 </tr>150 </tr>
151 <tr>151 <tr>
152 <td>workspaceSize(uint64_t*)</td>152 <td>workspaceSize(uint64_t*)</td>
@@ -98,7 +98,7 @@ aclnnStatus aclnnDequantBias(
98 <td>INT32</td>98 <td>INT32</td>
99 <td>ND</td>99 <td>ND</td>
100 <td>2</td>100 <td>2</td>
101- <td>×</td>101+ <td>√</td>
102 </tr>102 </tr>
103 <tr>103 <tr>
104 <td>weightScale (aclTensor*)</td>104 <td>weightScale (aclTensor*)</td>
@@ -108,7 +108,7 @@ aclnnStatus aclnnDequantBias(
108 <td>FLOAT, BFLOAT16</td>108 <td>FLOAT, BFLOAT16</td>
109 <td>ND</td>109 <td>ND</td>
110 <td>1</td>110 <td>1</td>
111- <td>×</td>111+ <td>√</td>
112 </tr>112 </tr>
113 <tr>113 <tr>
114 <td>activateScaleOptional (aclTensor*)</td>114 <td>activateScaleOptional (aclTensor*)</td>
@@ -118,7 +118,7 @@ aclnnStatus aclnnDequantBias(
118 <td>FLOAT</td>118 <td>FLOAT</td>
119 <td>ND</td>119 <td>ND</td>
120 <td>1</td>120 <td>1</td>
121- <td>×</td>121+ <td>√</td>
122 </tr>122 </tr>
123 <tr>123 <tr>
124 <td>biasOptional (aclTensor*)</td>124 <td>biasOptional (aclTensor*)</td>
@@ -128,7 +128,7 @@ aclnnStatus aclnnDequantBias(
128 <td>FLOAT, BFLOAT16, FLOAT16, INT32</td>128 <td>FLOAT, BFLOAT16, FLOAT16, INT32</td>
129 <td>ND</td>129 <td>ND</td>
130 <td>1</td>130 <td>1</td>
131- <td>×</td>131+ <td>√</td>
132 </tr>132 </tr>
133 <tr>133 <tr>
134 <td>outputDtype (int64_t)</td>134 <td>outputDtype (int64_t)</td>
@@ -148,7 +148,7 @@ aclnnStatus aclnnDequantBias(
148 <td>FLOAT16, BFLOAT16</td>148 <td>FLOAT16, BFLOAT16</td>
149 <td>ND</td>149 <td>ND</td>
150 <td>2</td>150 <td>2</td>
151- <td>×</td>151+ <td>√</td>
152 </tr>152 </tr>
153 <tr>153 <tr>
154 <td>workspaceSize (uint64_t*)</td>154 <td>workspaceSize (uint64_t*)</td>
@@ -9,4 +9,4 @@
9 9 
10message(STATUS "=== Debug: start dequant.bias.op_host.CMakeLists.txt ")10message(STATUS "=== Debug: start dequant.bias.op_host.CMakeLists.txt ")
11 11 
12-add_modules_sources(HOSTNAME ${OPHOST_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR} OPTYPE dequant_bias ACLNNTYPE aclnn)12+add_modules_sources(HOSTNAME ${OPHOST_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR} OPTYPE dequant_bias ACLNNTYPE aclnn_exclude)
@@ -28,7 +28,8 @@ public:
28 .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,28 .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
29 ge::FORMAT_ND, ge::FORMAT_ND})29 ge::FORMAT_ND, ge::FORMAT_ND})
30 .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,30 .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
31- ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});31+ ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
32+ .AutoContiguous();
atomgit-bot
atomgit-botatomgit-bot8 days ago

🟡 Medium Priority

改动:dequant_bias_def.cpp:32/41/50/59/68 为 x、weight_scale、activate_scale、bias、y 全部加上 .AutoContiguous(),即 aclnn 生成层不再拒绝非连续 Tensor,行为契约由「不支持」变为「支持」。 证据链:本仓 quant/dequant_bias/docs/aclnnDequantBias.md、docs/aclnnDequantBias_en.md 存在,而关联 issue #8693 明确记载该参数表五个参数的「非连续Tensor」列全部为 ×;本次 diff 只包含 op_host/dequant_bias_def.cpp 一个文件,PR 描述中却声称「文档更新:docs/aclnnDequantBias.md、aclnnDequantBias_en.md(非连续Tensor 列 ×→√,5 行×2)」——该文档改动未出现在本次变更中。 失败模式:合入后代码接受非连续输入/出参,资料仍声明 ×(不支持)。用户与下游 ST 依据资料构造用例时会认为该路径应返回 161002,而实际返回 0;反向地,若后续有人依据资料 × 再补「拒绝」校验,会与本次 AutoContiguous 直接冲突。这是本次单侧改动(只改代码不改契约文档)新引入的不一致。 修复方向:把 docs 两个文件的「非连续Tensor」列 5 行 ×→√ 一并纳入本 PR(或拆单时在描述中明确拆出并标注依赖),确保代码与资料同一提交内一致。

建议:将 quant/dequant_bias/docs/aclnnDequantBias.md 与 aclnnDequantBias_en.md 中 5 个参数的「非连续Tensor」列由 × 改为 √,与本 def.cpp 改动放在同一次提交中。

likedislike
Incorrect?
32 this->Input("weight_scale")33 this->Input("weight_scale")
33 .ParamType(REQUIRED)34 .ParamType(REQUIRED)
34 .DataType({ge::DT_BF16, ge::DT_BF16, ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT,35 .DataType({ge::DT_BF16, ge::DT_BF16, ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT,
@@ -36,7 +37,8 @@ public:
36 .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,37 .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
37 ge::FORMAT_ND, ge::FORMAT_ND})38 ge::FORMAT_ND, ge::FORMAT_ND})
38 .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,39 .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
39- ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});40+ ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
41+ .AutoContiguous();
40 this->Input("activate_scale")42 this->Input("activate_scale")
41 .ParamType(OPTIONAL)43 .ParamType(OPTIONAL)
42 .DataType({ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT,44 .DataType({ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT,
@@ -44,7 +46,8 @@ public:
44 .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,46 .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
45 ge::FORMAT_ND, ge::FORMAT_ND})47 ge::FORMAT_ND, ge::FORMAT_ND})
46 .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,48 .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
47- ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});49+ ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
50+ .AutoContiguous();
48 this->Input("bias")51 this->Input("bias")
49 .ParamType(OPTIONAL)52 .ParamType(OPTIONAL)
50 .DataType({ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT, ge::DT_INT32, ge::DT_BF16, ge::DT_FLOAT16,53 .DataType({ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT, ge::DT_INT32, ge::DT_BF16, ge::DT_FLOAT16,
@@ -52,7 +55,8 @@ public:
52 .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,55 .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
53 ge::FORMAT_ND, ge::FORMAT_ND})56 ge::FORMAT_ND, ge::FORMAT_ND})
54 .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,57 .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
55- ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});58+ ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
59+ .AutoContiguous();
56 this->Output("y")60 this->Output("y")
57 .ParamType(REQUIRED)61 .ParamType(REQUIRED)
58 .DataType({ge::DT_BF16, ge::DT_BF16, ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16,62 .DataType({ge::DT_BF16, ge::DT_BF16, ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16,
@@ -60,7 +64,8 @@ public:
60 .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,64 .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
61 ge::FORMAT_ND, ge::FORMAT_ND})65 ge::FORMAT_ND, ge::FORMAT_ND})
62 .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,66 .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
63- ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});67+ ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
68+ .AutoContiguous();
atomgit-bot
atomgit-botatomgit-bot8 days ago

🟡 Medium Priority

改动:dequant_bias_def.cpp:68 在 Output("y") 上新增 .AutoContiguous()(32/41/50 行同理作用于输入)。 证据链:

  1. 关联 issue #8663 给出的实测结论是:没有手写 op_host/op_api/ 的算子出参非连续全部静默错值(7/7 失败),有手写实现(aclnnHardswish、aclnnForeachMulList)才正常;根因是自动生成层缺「算到连续 buffer 后按 view 回写」。quant/dequant_bias 目录下确实没有 op_host/op_api/(仅 dequant_bias_def.cpp / _infershape.cpp / _tiling.cpp / _tiling_arch35.cpp),与本 PR 描述的「缺口在生成层」自述一致。
  2. 同目录 dequant_bias_tiling_arch35.cpp:253-263 存在未改动的强校验:yShape.GetStorageShape() 必须等于 x 的 [M, N]。若框架的连续化发生在 tiling 之后(oracle 输入仍是用户视图 storage=[40,512]),非连续出参在 950/350 平台会直接 GRAPH_FAILED,与本次同步宣称的「支持非连续」相反;若连续化发生在 tiling 之前,则 kernel 只写连续 buffer,必须由生成层做 view 回写,而这正是 issue #8663 认定缺失的一环。 失败模式:出参 y 传 stride 带空洞的 view 时,要么仍按连续行主序写满 storage 前缀(一半位置留旧值、两段返回 0),要么在 arch35 路径被 tiling 校验拒绝——两者都无法达到 PR 宣称的「按 stride 正确写入/回读」,而 PR 自己把上板验证交由 CI 且无任何本地证据。 修复方向:在合入前用 #8693 的复现脚本([40,256]/{512,2}/[40,512])在 910b 与 950 各跑一次,确认按 stride 读回逐元素等于基线;若框架并未在 aclnn 单算子路径支持出参回写,则应改为在 host 侧显式拒绝(161002)并在资料保留 ×,或补手写 op_host/op_api/ 做 view copy(参照 foreach 族 ForeachViewCopyToOutputTensorList)。

建议:合入前用 #8693 复现脚本在 ascend910b 与 ascend950 验证非连续出参按 stride 读回与基线逐元素一致;若自动生成层不做出参 view 回写,则改为 host 侧显式返回 161002 并保留资料 ×,或补手写 op_host/op_api 实现连续化后回写。

likedislike
Incorrect?
64 this->Attr("output_dtype").AttrType(REQUIRED).Int(); // 输出类型,传入69 this->Attr("output_dtype").AttrType(REQUIRED).Int(); // 输出类型,传入
65 this->AICore().AddConfig("ascend910b");70 this->AICore().AddConfig("ascend910b");
66 this->AICore().AddConfig("ascend910_93");71 this->AICore().AddConfig("ascend910_93");
@@ -0,0 +1,151 @@
1+/**
2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+#include "aclnn_dequant_bias.h"
11+#include "dequant_bias.h"
12+#include "aclnn_kernels/contiguous.h"
13+#include "aclnn_kernels/common/op_error_check.h"
14+#include "aclnn/aclnn_base.h"
15+#include "opdev/common_types.h"
16+#include "opdev/format_utils.h"
17+#include "opdev/shape_utils.h"
18+#include "opdev/op_dfx.h"
19+#include "opdev/op_executor.h"
20+#include "opdev/op_log.h"
21+#include "opdev/data_type_utils.h"
22+#include "opdev/tensor_view_utils.h"
23+#include "opdev/make_op_executor.h"
24+ 
25+using namespace op;
26+#ifdef __cplusplus
27+extern "C" {
28+#endif
29+ 
30+static const std::initializer_list<op::DataType> DTYPE_SUPPORT_X_LIST = {op::DataType::DT_INT32};
31+static const std::initializer_list<op::DataType> DTYPE_SUPPORT_WS_LIST = {op::DataType::DT_FLOAT,
32+ op::DataType::DT_BF16};
33+static const std::initializer_list<op::DataType> DTYPE_SUPPORT_AS_LIST = {op::DataType::DT_FLOAT};
34+static const std::initializer_list<op::DataType> DTYPE_SUPPORT_BIAS_LIST = {
35+ op::DataType::DT_FLOAT, op::DataType::DT_BF16, op::DataType::DT_FLOAT16, op::DataType::DT_INT32};
36+static const std::initializer_list<op::DataType> DTYPE_SUPPORT_OUT_LIST = {op::DataType::DT_FLOAT16,
37+ op::DataType::DT_BF16};
38+ 
39+static bool CheckNotNull(const aclTensor* x, const aclTensor* weightScale, const aclTensor* activateScale,
40+ const aclTensor* bias, const aclTensor* out)
41+{
42+ OP_CHECK_NULL(x, return false);
43+ OP_CHECK_NULL(weightScale, return false);
44+ OP_CHECK_NULL(out, return false);
45+ return true;
46+}
47+ 
48+static bool CheckDtypeValid(const aclTensor* x, const aclTensor* weightScale, const aclTensor* activateScale,
49+ const aclTensor* bias, const aclTensor* out)
50+{
51+ OP_CHECK_DTYPE_NOT_SUPPORT(x, DTYPE_SUPPORT_X_LIST, return false);
52+ OP_CHECK_DTYPE_NOT_SUPPORT(weightScale, DTYPE_SUPPORT_WS_LIST, return false);
53+ if (activateScale != nullptr) {
54+ OP_CHECK_DTYPE_NOT_SUPPORT(activateScale, DTYPE_SUPPORT_AS_LIST, return false);
55+ }
56+ if (bias != nullptr) {
57+ OP_CHECK_DTYPE_NOT_SUPPORT(bias, DTYPE_SUPPORT_BIAS_LIST, return false);
58+ }
59+ OP_CHECK_DTYPE_NOT_SUPPORT(out, DTYPE_SUPPORT_OUT_LIST, return false);
60+ return true;
61+}
62+ 
63+static bool CheckOutputDtypeValid(int64_t outputDtype)
64+{
65+ OP_CHECK(
66+ outputDtype == 1 || outputDtype == 27,
67+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "outputDtype should be 1 (FLOAT16) or 27 (BFLOAT16), got %ld.", outputDtype),
68+ return false);
69+ return true;
70+}
71+ 
72+static bool CheckShape(const aclTensor* x, const aclTensor* out)
73+{
74+ // out 的 view shape 必须与 x 一致(kernel 按 x 的 [M, N] 生成输出)
75+ OP_CHECK_SHAPE_NOT_EQUAL(x, out, return false);
76+ return true;
77+}
78+ 
79+static aclnnStatus CheckParams(const aclTensor* x, const aclTensor* weightScale, const aclTensor* activateScale,
80+ const aclTensor* bias, int64_t outputDtype, const aclTensor* out)
81+{
82+ CHECK_RET(CheckNotNull(x, weightScale, activateScale, bias, out), ACLNN_ERR_PARAM_NULLPTR);
83+ CHECK_RET(CheckDtypeValid(x, weightScale, activateScale, bias, out), ACLNN_ERR_PARAM_INVALID);
84+ CHECK_RET(CheckOutputDtypeValid(outputDtype), ACLNN_ERR_PARAM_INVALID);
85+ CHECK_RET(CheckShape(x, out), ACLNN_ERR_PARAM_INVALID);
86+ return ACLNN_SUCCESS;
87+}
88+ 
89+aclnnStatus aclnnDequantBiasGetWorkspaceSize(const aclTensor* x, const aclTensor* weightScale,
90+ const aclTensor* activateScaleOptional, const aclTensor* biasOptional,
91+ int64_t outputDtype, const aclTensor* out, uint64_t* workspaceSize,
92+ aclOpExecutor** executor)
93+{
94+ OP_CHECK_COMM_INPUT(workspaceSize, executor);
95+ 
96+ L2_DFX_PHASE_1(aclnnDequantBias, DFX_IN(x, weightScale, activateScaleOptional, biasOptional), DFX_OUT(out));
97+ 
98+ auto uniqueExecutor = CREATE_EXECUTOR();
99+ CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);
100+ 
101+ auto ret = CheckParams(x, weightScale, activateScaleOptional, biasOptional, outputDtype, out);
102+ CHECK_RET(ret == ACLNN_SUCCESS, ret);
103+ 
104+ auto xContiguous = l0op::Contiguous(x, uniqueExecutor.get());
105+ CHECK_RET(xContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
106+ auto weightScaleContiguous = l0op::Contiguous(weightScale, uniqueExecutor.get());
107+ CHECK_RET(weightScaleContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
108+ const aclTensor* activateScaleContiguous = nullptr;
109+ if (activateScaleOptional != nullptr) {
110+ activateScaleContiguous = l0op::Contiguous(activateScaleOptional, uniqueExecutor.get());
111+ CHECK_RET(activateScaleContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
112+ }
113+ const aclTensor* biasContiguous = nullptr;
114+ if (biasOptional != nullptr) {
115+ biasContiguous = l0op::Contiguous(biasOptional, uniqueExecutor.get());
116+ CHECK_RET(biasContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
117+ }
118+ 
119+ // 输出out连续且 storage 与 view 一致、dtype 与 outputDtype 推导一致时,
120+ // 直接绑定 out 作为 kernel 输出(kernel 直接写入调用方内存);
121+ // 非连续时先算到连续张量,再按 out 的 stride 回写
122+ const bool directOut = IsContiguous(out) && out->GetStorageShape() == out->GetViewShape() &&
wangzitao
wangzitaowangzitao7 days ago

[Major] 新增的 directOut 直接绑定只校验了 out 自身连续、storage == view 与 dtype,未校验 out 的 view shape 与 x(即 kernel 按 x->GetViewShape() 生成的输出形状)一致。CheckParams 目前只做非空/dtype 校验(本文件 :82-89),没有输出 shape 校验;旧路径至少经过 l0op::ViewCopy,直接路径则由 kernel 按 x 的 tiling 写入调用方内存。调用方若传入 out 为形状更小的连续张量(例如 x 为 [8]、out 为 [4],storage==view 且 FLOAT16),kernel 会按 8 个元素写入 4 元素缓冲区,造成越界写;形状更大时则静默返回错误结果。建议在 directOut 判定前增加 out->GetViewShape() == x->GetViewShape()(或等价 shape 校验,失败返回 ACLNN_ERR_PARAM_INVALID);仓内同类直接写 out 的实现(如 foreach/foreach_mul_scalar/op_host/op_api/aclnn_foreach_mul_scalar_v2.cpp:76-90)均先做 OP_CHECK_SHAPE_NOT_EQUAL。

agent检视,通过wangzitao_leo提交

likedislike
123+ ((outputDtype == 1 && out->GetDataType() == op::DataType::DT_FLOAT16) ||
124+ (outputDtype == 27 && out->GetDataType() == op::DataType::DT_BF16));
125+ const aclTensor* dequantBiasOut = nullptr;
126+ if (directOut) {
127+ dequantBiasOut = l0op::DequantBias(xContiguous, weightScaleContiguous, activateScaleContiguous, biasContiguous,
128+ outputDtype, uniqueExecutor.get(), out);
129+ CHECK_RET(dequantBiasOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
130+ } else {
131+ dequantBiasOut = l0op::DequantBias(xContiguous, weightScaleContiguous, activateScaleContiguous, biasContiguous,
132+ outputDtype, uniqueExecutor.get());
133+ CHECK_RET(dequantBiasOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
134+ auto viewCopyResult = l0op::ViewCopy(dequantBiasOut, out, uniqueExecutor.get());
135+ CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
136+ }
137+ 
138+ *workspaceSize = uniqueExecutor->GetWorkspaceSize();
139+ uniqueExecutor.ReleaseTo(executor);
140+ return ACLNN_SUCCESS;
141+}
142+ 
143+aclnnStatus aclnnDequantBias(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, const aclrtStream stream)
144+{
145+ L2_DFX_PHASE_2(aclnnDequantBias);
146+ return CommonOpExecutorRun(workspace, workspaceSize, executor, stream);
147+}
148+ 
149+#ifdef __cplusplus
150+}
151+#endif
@@ -0,0 +1,85 @@
1+/**
2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+#ifndef OP_API_INC_ACLNN_DEQUANT_BIAS_H_
11+#define OP_API_INC_ACLNN_DEQUANT_BIAS_H_
12+ 
13+#include "aclnn/aclnn_base.h"
14+#include "aclnn_util.h"
15+ 
16+#ifdef __cplusplus
17+extern "C" {
18+#endif
19+ 
20+/**
21+ * @brief aclnnDequantBias的第一段接口,根据具体的计算流程,计算workspace大小。
22+ * @domain aclnn_ops_infer
23+ *
24+ * 算子功能:反量化偏置叠加:
25+ * $$
26+ * y = A \times weightScale \times activateScale
27+ * $$
28+ * (无 bias)
29+ * $$
30+ * y = (A + bias) \times weightScale \times activateScale
johan_lee
johan_leejohan_lee8 days ago

【review】 问题分类:正确性 位置:quant/dequant_bias/op_host/op_api/aclnn_dequant_bias.h:26,代码:* y = (A + bias) \times weightScale \times activateScale 问题:新头文件把 bias 一律写成反量化前加。实际 INT32 bias 在乘缩放因子前加,FLOAT、FLOAT16 和 BFLOAT16 bias 在两次缩放后加。arch22 的 dequant_bias_impl.h:232、:244 和 arch35 的 dequant_bias_kernel.h:305、:320 都遵循这个顺序。例如 A=2、FLOAT bias=1、weightScale=2、activateScale=1 时,实际输出 5,按新注释会得到 6。 可能导致的结果:按公开头文件编写参考计算或选择 bias 数据类型时,浮点 bias 在缩放因子非 1 的场景会产生错误预期。 建议的修改方法:在头文件中分别列出无 bias、INT32 bias 和浮点 bias 三种公式,并注明缺省 activateScale 按 1 处理。

likedislike
31+ * $$
32+ * (bias 为 INT32 时,先加 bias 再缩放)
33+ * $$
34+ * y = A \times weightScale \times activateScale + bias
35+ * $$
36+ * (bias 为 FLOAT、FLOAT16、BFLOAT16 时,先缩放再加 bias;
37+ * activateScale 缺省时按 1 处理)
38+ *
39+ * 实现说明:
40+ * api计算的基本路径:
41+ * ```mermaid
42+ * graph LR
43+ * A[(x, weightScale, activateScale, bias)] -->B([L0::Contiguous])
44+ * B --> C([L0::DequantBias])
45+ * C --> D([L0::ViewCopy])
46+ * D --> E[(out)]
47+ * ```
48+ *
49+ * @param [in] x: device侧的aclTensor,shape为[M, N],数据类型支持INT32,支持非连续的Tensor。
50+ * @param [in] weightScale: device侧的aclTensor,shape为[N],长度与x的N维长度一致,
51+ * 数据类型支持FLOAT、BFLOAT16,支持非连续的Tensor。
52+ * @param [in] activateScaleOptional: device侧的aclTensor,shape为[M],长度与x的M维长度一致,
53+ * 数据类型支持FLOAT,支持非连续的Tensor。可选输入,传入nullptr时视为不参与计算。
54+ * @param [in] biasOptional: device侧的aclTensor,shape为[N],长度与x的N维长度一致,
55+ * 数据类型支持FLOAT、BFLOAT16、FLOAT16、INT32,支持非连续的Tensor。可选输入,传入nullptr时视为不参与计算。
56+ * @param [in] outputDtype: 表示输出out的数据类型,值为1表示FLOAT16,值为27表示BFLOAT16。
57+ * 当weightScale数据类型为FLOAT时该参数配置为1,为BFLOAT16时配置为27。
58+ * @param [in] out: device侧的aclTensor,shape为[M, N],数据类型支持FLOAT16、BFLOAT16,支持非连续的Tensor。
59+ * @param [out] workspaceSize: 返回用户需要在npu device侧申请的workspace大小。
60+ * @param [out] executor: 返回op执行器,包含算子计算流程。
61+ * @return aclnnStatus: 返回状态码。
62+ */
63+ACLNN_API aclnnStatus aclnnDequantBiasGetWorkspaceSize(const aclTensor* x, const aclTensor* weightScale,
64+ const aclTensor* activateScaleOptional,
65+ const aclTensor* biasOptional, int64_t outputDtype,
66+ const aclTensor* out, uint64_t* workspaceSize,
67+ aclOpExecutor** executor);
68+ 
69+/**
70+ * @brief aclnnDequantBias的第二段接口,用于执行计算。
71+ *
72+ * @param [in] workspace: 在npu device侧申请的workspace内存起址。
73+ * @param [in] workspaceSize: 在npu device侧申请的workspace大小,由第一段接口获取。
74+ * @param [in] executor: op执行器,包含了算子计算流程。
75+ * @param [in] stream: acl stream流。
76+ * @return aclnnStatus: 返回状态码。
77+ */
78+ACLNN_API aclnnStatus aclnnDequantBias(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor,
79+ const aclrtStream stream);
80+ 
81+#ifdef __cplusplus
82+}
83+#endif
84+ 
85+#endif // OP_API_INC_ACLNN_DEQUANT_BIAS_H_
@@ -0,0 +1,44 @@
1+/**
2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+#include "dequant_bias.h"
11+#include "opdev/make_op_executor.h"
12+#include "opdev/op_def.h"
13+#include "opdev/op_dfx.h"
14+#include "opdev/op_log.h"
15+ 
16+using namespace op;
17+ 
18+namespace l0op {
19+ 
20+OP_TYPE_REGISTER(DequantBias);
21+ 
22+const aclTensor* DequantBias(const aclTensor* x, const aclTensor* weightScale, const aclTensor* activateScale,
23+ const aclTensor* bias, int64_t outputDtype, aclOpExecutor* executor, const aclTensor* out)
24+{
25+ L0_DFX(DequantBias, x, weightScale, activateScale, bias, outputDtype);
26+ DataType outputDataType = (outputDtype == 1) ? DataType::DT_FLOAT16 : DataType::DT_BF16;
27+ const aclTensor* outTensor = out;
28+ if (outTensor == nullptr) {
29+ auto outputShape = x->GetViewShape();
30+ outTensor = executor->AllocTensor(outputShape, outputDataType, Format::FORMAT_ND);
31+ if (outTensor == nullptr) {
32+ return nullptr;
33+ }
34+ }
35+ auto ret = ADD_TO_LAUNCHER_LIST_AICORE(DequantBias, OP_INPUT(x, weightScale, activateScale, bias),
36+ OP_OUTPUT(outTensor), OP_ATTR(outputDtype));
37+ if (ret != ACLNN_SUCCESS) {
38+ OP_LOGE(ACLNN_ERR_INNER_STATIC_WORKSPACE_INVALID, "DequantBias launch failed.");
39+ return nullptr;
40+ }
41+ return outTensor;
42+}
43+ 
44+} // namespace l0op
@@ -0,0 +1,23 @@
1+/**
2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+#ifndef OP_API_DEQUANT_BIAS_H_
11+#define OP_API_DEQUANT_BIAS_H_
12+ 
13+#include "opdev/op_executor.h"
14+ 
15+namespace l0op {
16+ 
17+const aclTensor* DequantBias(const aclTensor* x, const aclTensor* weightScale, const aclTensor* activateScale,
18+ const aclTensor* bias, int64_t outputDtype, aclOpExecutor* executor,
19+ const aclTensor* out = nullptr);
20+ 
21+} // namespace l0op
22+ 
23+#endif // OP_API_DEQUANT_BIAS_H_